Flink Forward 2015

1. A Comparative Performance
Evaluation of Flink
Dongwon Kim
POSTECH

2. About Me
• Postdoctoral researcher @ POSTECH
• Research interest
• Design and implementation of distributed systems
• Performance optimization of big data processing engines
• Doctoral thesis
• MR2: Fault Tolerant MapReduce with the Push Model
• Personal blog
• http://eastcirclek.blogspot.kr
• Why I’m here 
2

3. Outline
• TeraSort for various engines
• Experimental setup
• Results & analysis
• What else for better performance?
• Conclusion
3

4. TeraSort
• Hadoop MapReduce program for the annual terabyte sort competition
• TeraSort is essentially distributed sort (DS)
a4
b3 a1
a2
b1
b2
a2
b1 a3
a4
b3
b4
Disk
a2
a4
b3
b1
a1
b4
a3
b2
a1
a3
b4
b2
Disk
a2
a4
a1
a3
b3
b1
b4
b2
read shufflinglocal sort write
Disk
Disk
local sort
Node 1
Node 2
Typical DS phases :
4Total order

5. • Included in Hadoop distributions
• with TeraGen & TeraValidate
• Identity map & reduce functions
• Range partitioner built on sampling
• To guarantee a total order & to prevent partition skew
• Sampling to compute boundary points within few seconds
TeraSort for MapReduce
Reduce taskMap task
read shuffling sortsort reducemap write
read shufflinglocal sort writelocal sortDS phases :
reducemap
5
Record range
…
Partition 1 Partition 2 Partition r
boundary points

6. • Tez can execute TeraSort for MapReduce w/o any modification
• mapreduce.framework.name = yarn-tez
• Tez DAG plan of TeraSort for MapReduce
TeraSort for Tez
finalreduce vertex
initialmap vertex
Map task
read sortmap
Reduce task
shuffling sort reduce write
input data
output data
6

7. TeraSort for Spark & Flink
• My source code in GitHub:
• https://github.com/eastcirclek/terasort
• Sampling-based range partitioner from TeraSort for MapReduce
• Visit my personal blog for a detailed explanation
• http://eastcirclek.blogspot.kr
7

8. RDD1 RDD2
• Code
• Two RDDs
TeraSort for Spark
Stage 1Stage 0
Shuffle-Map Task
(for newAPIHadoopFile)
read sort
Result Task
(for repartitionAndSortWithinPartitions)
shuffling sort write
read shufflinglocal sort writelocal sort
Create a new RDD to read from HDFS
# partitions = # blocks
Repartition the parent RDD
based on the user-specified partitioner
Write output to HDFS
DS phases :
8

9. Pipeline
• Code
• Pipelines consisting of four operators
TeraSort for Flink
read shuffling writelocal sort
Create a dataset to read tuples
from HDFS
partition tuples
Sort tuples of each partition
DataSource Partition SortPartition DataSink
local sort
No map-side sorting
due to pipelined execution
Write output to HDFS
DS phases :
9

10. Importance of TeraSort
• Suitable for measuring the pure performance of big data engines
• No data transformation (like map, filter) with user-defined logic
• Basic facilities of each engine are used
• “Winning the sort benchmark” is a great means of PR
10

11. Outline
• TeraSort for various engines
• Experimental setup
• Machine specification
• Node configuration
• Results & analysis
• What else for better performance?
• Conclusion
11

12. Machine specification (42 identical machines)
DELL PowerEdge R610
CPU
Two X5650 processors
(Total 12 cores)
Memory
Total 24Gb
Disk
6 disks * 500GB/disk
Network
10 Gigabit Ethernet
My machine Spark team
Processor
Intel Xeon X5650
(Q1, 2010)
Intel Xeon E5-2670
(Q1, 2012)
Cores 6 * 2 processors 8 * 4 processors
Memory 24GB 244GB
Disks 6 HDD's 8 SSD's
Results can be different
in newer machines
12

13. 24GB on each node
Node configuration
Total 2 GB
for daemons
13 GB
Tez-0.7.0
NodeManager (1 GB)
ShuffleService
MapTask (1GB)
DataNode (1 GB)
MapTask (1GB)
ReduceTask (1GB)
ReduceTask (1GB)
MapTask (1GB)
MapTask (1GB)
MapTask (1GB)
…
MapReduce-2.7.1
NodeManager (1 GB)
ShuffleService
DataNode (1 GB)
MapTask (1GB)
MapTask (1GB)
ReduceTask (1GB)
ReduceTask (1GB)
MapTask (1GB)
MapTask (1GB)
MapTask (1GB)
…
12 GB
Flink
Spark
Spark-1.5.1
NodeManager (1 GB)
Executor (12GB)
Internal
memory layout
Various managers
DataNode (1 GB)
Task slot 1
Task slot 2
Task slot 12
...
Thread pool
Flink-0.9.1
NodeManager (1 GB)
TaskManager (12GB)
DataNode (1 GB)
Internal
memory layout
Various managers
Task slot 1
Task slot 2
Task slot 12
...
Task threads
Tez
MapReduce
ReduceTask (1GB)ReduceTask (1GB)
13
12 simultaneous tasks
at most
Driver (1GB) JobManager (1GB)

14. Outline
• TeraSort for various engines
• Experimental setup
• Results & analysis
• Flink is faster than other engines due to its pipelined execution
• What else for better performance?
• Conclusion
14

15. How to read a swimlane graph & throughput graphs
Tasks
Time since job starts (seconds)
2nd stage
1st
2nd
3rd
4th
5th
6th
15
Cluster network throughput
Cluster disk throughput
In
Out
Disk read
Disk
Write
- 6 waves of 1st stage tasks
- 1 wave of 2nd stage tasks
- Two stages are hardly overlapped
1st stage
2nd stage
1st stage
2nd stage
No network traffic during 1st stage
Each line : duration of each task
Different patterns for different stages

16. Result of sorting 80GB/node (3.2TB)
1480 sec
1st stage
1st stage
1st stage
2nd stage
2157 sec
2nd stage
2171 sec
1 DataSource
2 Partition
3 SortPartition
4 DataSink
• Flink is the fastest due to its pipelined execution
• Tez and Spark do not overlap 1st and 2nd stages
• MapReduce is slow despite overlapping stages
MapReduce
in Hadoop-2.7.1
Tez-0.7.0
Spark-1.5.1
Flink-0.9.1
2nd stage
1887 sec
2157
1887
2171
1480
0
500
1000
1500
2000
2500
MapReduce
in Hadoop-2.7.1
Tez-0.7.0 Spark-1.5.1 Flink-0.9.1
Time(seconds)
16* Map output compression turned on for Spark and Tez
* *

17. Tez and Spark do not overlap 1st and 2nd stages
Cluster network
throughput
Cluster disk throughput
In
Out
Disk read
Cluster network
throughput
Cluster disk throughput
In
Out
Disk read
Disk
write
Disk
write
Disk read
Disk write
Out
In
(1) 2nd stage starts
(2)
Output of 1st stage is sent
(1) 2nd stage starts
(2)
Output of 1st stage is sent
(1)
Network traffic
occurs from start
Cluster network
throughput
(2)
Write to HDFS occurs
right after shuffling is done
1 DataSource
2 Partition
3 SortPartition
4 DataSink
idle idle
(3)
Disk write to HDFS occurs
after shuffling is done
(3)
Disk write to HDFS occurs
after shuffling is done
17

18. Tez does not overlap 1st and 2nd stages
• Tez has parameters to control the degree of overlap
• tez.shuffle-vertex-manager.min-src-fraction : 0.2
• tez.shuffle-vertex-manager.max-src-fraction : 0.4
• However, 2nd stage is scheduled early but launched late
scheduled launched
18

19. Spark does not overlap 1st and 2nd stages
• Spark cannot execute multiple stages simultaneously
• also mentioned in the following VLDB paper (2015)
Spark doesn’t support the overlap
between shuffle write and read stages.
…
Spark may want to support this overlap
in the future to improve performance.
Experimental results of this paper
- Spark is faster than MapReduce for WordCount, K-means, PageRank.
- MapReduce is faster than Spark for Sort.
19

20. MapReduce is slow despite overlapping stages
• mapreduce.job.reduce.slowstart.completedMaps : [0.0, 1.0]
• Wang’s attempt to overlap spark stages
0.05
(overlapping, default)
0.95
(no overlapping)
2157 sec
10%
improvement
20
Wang proposes to overlap stages
to achieve better utilization
10%???
Why Spark & MapReduce
improve just 10%?
2385 sec
2nd stage
1st stage

21. Disk
Data transfer between tasks of different stages
Output file
P1 P2 Pn
Shuffle server
…
Consumer
Task 1
Consumer
Task 2
Consumer
Task n
P1
…
P2
Pn
Traditional pull model
- Used in MapReduce, Spark, Tez
- Extra disk access & simultaneous disk access
- Shuffling affects the performance of producers
Producer
Task
(1)
Write output
to disk
(2)
Request P1
(3)
Send P1
Pipelined data transfer
- Used in Flink
- Data transfer from memory to memory
- Flink causes fewer disk access during shuffling
21
Leads to only 10% improvement

22. Flink causes fewer disk access during shuffling
Map
Reduce
Flink diff.
Total disk write
(TB)
9.9 6.5 3.4
Total disk read
(TB)
8.1 6.9 1.2
Difference comes
from shuffling
Shuffled data are sometimes
read from page cache
Cluster disk throughput
Disk read
Disk write
Disk read
Disk write
Cluster disk throughput
FlinkMapReduce
22
Total amount of disk read/write
equals to
the area of blue/green region

23. Result of TeraSort with various data sizes
node data size
(GB)
Time (seconds)
Flink Spark MapReduce Tez
10 157 387 259 277
20 350 652 555 729
40 741 1135 1085 1709
80 1480 2171 2157 1887
160 3127 4927 4796 3950
23
100
1000
10000
10 20 40 80 160
Time(seconds)
node data size (GB)
Flink Spark MapReduce Tez
What we’ve seen
Log scale
* Map output compression turned on for Spark and Tez

24. Result of HashJoin
• 10 slave nodes
• org.apache.tez.examples.JoinDataGen
• Small dataset : 256MB
• Large dataset : 240GB (24GB/node)
• Result :
24
Visit my blog

Flink is
~2x faster than Tez
~4x faster than Spark
770
1538
378
0
200
400
600
800
1000
1200
1400
1600
1800
Tez-0.7.0 Spark-1.5.1 Flink-0.9.1
Time(seconds)
* No map output compression for both Spark and Tez unlike in TeraSort

25. Result of HashJoin with swimlane & throughput graphs
25
Idle
1 DataSource
2 DataSource
3 Join
4 DataSink
Idle
Cluster network throughput
Cluster disk throughput
In
Out
Disk
read
Disk
write
Disk read
Disk write
In
Out
In
Out
Disk read
Disk
write
Cluster network throughput
Cluster disk throughput
0.24 TB
0.41 TB
0.60 TB 0.84 TB
0.68 TB
0.74 TB
Overlap
2nd
3rd

26. Flink’s shortcoming
• No support for map output compression
• Small data blocks are pipelined between operators
• Job-level fault tolerance
• Shuffle data are not materialized
• Low disk throughput during the post-shuffling phase
26

27. Low disk throughput during the post-shuffling phase
• Possible reason : sorting records from small files
• Concurrent disk access to small files  too many disk seeks
 low disk throughput
• Other engines merge records from larger files than Flink
• “Eager pipelining moves some of the sorting work from the mapper to the
reducer”
• from MapReduce online (NSDI 2010)
Flink Tez MapReduce
27

28. Outline
• TeraSort for various engines
• Experimental setup
• Results & analysis
• What else for better performance?
• Conclusion
28

29. MR2 – another MapReduce engine
• PhD thesis
• MR2: Fault Tolerant MapReduce with the Push Model
• developed for 3 years
• Provide the user interface of Hadoop MapReduce
• No DAG support
• No in-memory computation
• No iterative-computation
• Characteristics
• Push model + Fault tolerance
• Techniques to boost up HDD throughput
• Prefetching for mappers
• Preloading for reducers
29

30. MR2 pipeline
• 7 types of components with memory buffers
1. Mappers & reducers : to apply user-defined functions
2. Prefetcher & preloader : to eliminate concurrent disk access
3. Sender & reducer & merger : to implement MR2’s push model
• Various buffers : to pass data between components w/o disk IOs
• Minimum disk access (2 disk reads & 2 disk writes)
• +1 disk write for fault tolerance
W1 R2 W2R1
30
1 12 23 3 3
W3

31. Prefetcher & Mappers
• Prefetcher loads data for multiple mappers
• Mappers do not read input from disks
<MR2><Hadoop MapReduce>
Mapper1 processing Blk1
Mapper2 processing Blk2
Time
Disk
throughput
CPU
utilization
2 mappers
on a node
Blk1
Time
Prefetcher Blk2 Blk3
Blk1
2
Blk1
1
Blk2
2
Blk2
1
Blk3
2
Blk3
1
Blk4
Blk4
2
Blk4
1
Disk
throughput
CPU
utilization
2 mappers
on a node
31

32. Push-model in MR2
• Node-to-node network connection for pushing data
• To reduce # network connections
• Data transfer from memory buffer
• Mappers stores spills in send buffer
• Spills are pushed to reducer sides by sender
• Fault tolerance (can be turned on/off)
• Input ranges of each spill are known to master for reproduce
• For fast recovery
• store spills on disk for fast recovery (extra disk write)
32
similar to Flink’s pipelined execution
MR2 does local sorting
before pushing data
similar to Spark

33. Receiver’s managed memory
Receiver & merger & preloader & reducer
• Merger produces a file from different partition data
• sorts each partition data
• and then does interleaving
• Preloader preloads each group into reduce buffer
• Reducers do not read data directly from disks
• MR2 can eliminate concurrent disk reads from reducers thanks to Preloader
P1 P2 P3 P4
P1 P2 P3 P4
P1 P2 P3 P4
… …
Preloader loads each group
(1 disk access for 4 partitions)
33

34. Result of sorting 80GB/node (3.2TB) with MR2
MapReduce
in Hadoop-2.7.1
Tez-0.7.0 Spark-1.5.1 Flink-0.9.1 MR2
Time (sec) 2157 1887 2171 1480 890
MR2 speedup
over other engines
2.42 2.12 2.44 1.66 -
2157
1887
2171
1480
890
0
500
1000
1500
2000
2500
MapReduce
in Hadoop-2.7.1
Tez-0.7.0 Spark-1.5.1 Flink-0.9.1 MR2
Time(seconds)
34

35. Disk & network throughput
1. DataSource / Mapping
• Prefetcher is effective
• MR2 shows higher disk
throughput
2. Partition / Shuffling
• Records to shuffle are
generated faster from in MR2
3. DataSink / Reducing
• Preloader is effective
• Almost 2x throughput
Disk read
Disk write
Out
In
Cluster network throughput
Cluster disk throughput
Out
In
Disk read
Disk write
Flink MR2
1
1
2
2
3
3
35

36. • Experimental results using 10 nodes
PUMA (PUrdue MApreduce benchmarks suite)
36

37. Outline
• TeraSort for various engines
• Experimental setup
• Experimental results & analysis
• What else for better performance?
• Conclusion
37

38. Conclusion
• Pipelined execution for both batch and streaming processing
• Even better than other batch processing engines for
TeraSort & HashJoin
• Shortcomings due to pipelined execution
• No fine-grained fault tolerance
• No map output compression
• Low disk throughput during the post-shuffling phase
38

39. Thank you!
Any question?
39